Let your AI agent watch what your competitors change every week

What you get

Your competitors keep shipping new things, and you can only guess what's working for them. This guide shows you how to hand REM, an open-source competitor research tool, to your AI agent. Once a week, it checks a competitor's public marketing footprint: ads, social media, search pages, and website redesigns. It compares the results with last week and helps you pick the one move most worth testing next week. You don't write any code. Three copy-paste prompts are included below.

Set your expectations first: the tool collects the evidence, your AI agent turns it into a report, and you decide what to test next week. Public data can't show a competitor's real ad spend or conversion rates, so anything the report can't see should simply be marked "unknown."

The loop works like this: run one full check on a competitor, read the four report sections, and turn one finding into next week's experiment. A week later, run the same setup again to see what they added and what they stopped, then decide your next experiment.

We walked through how this tool breaks down a company in the case Your rivals' growth playbook, laid bare. The install details for the tool itself are on the skill page reverse-engineer-marketing. This guide focuses on turning it into a weekly habit.

Know what's free and what costs money

What it can checkExtra cost?Official sites, tracking code, sitemaps, historical snapshots, app stores, Hacker NewsFree, no keys neededMeta, Google, and LinkedIn ad libraries, social data, founders, and RedditNeeds a Scrape Creators key, about 25 to 50 credits per runWriting the reportCan use the Claude Code or Codex subscription you already have, no separate API needed

The project author says that running the same check again within 24 hours uses a cache, so it doesn't spend credits a second time.

Step 1: Run the demo first, then do your first check on a competitor

Have your AI agent install the tool, look at the demo report that comes with the project to get a feel for it, and then check your competitor. Copy this:

Help me install the open-source tool reverse-engineer-marketing (REM), which I'll use to check a competitor's public marketing signals once a week.
Repo: https://github.com/dansugc/reverse-engineer-marketing
1. Read the README first. If the package isn't on npm yet, install it from source as the README describes. Tell me what you're going to install before you do it.
2. Open the demo that comes with the tool, and explain in plain language what the report looks like.
3. Then check my competitor: [competitor website]. Don't use any keys yet. Only run the free parts.
4. For writing the report, use the AI agent subscription I'm already signed into on this computer. Don't ask me to get an API key.
5. Tell me which items are still "unknown," how many credits it would cost to fill them in, and ask whether I want to. If a key is needed, teach me how to enter it myself in the settings. Don't ask me to paste the key into the chat.

Why it's written this way: the README says one command is enough, but the package wasn't on npm when we wrote this guide. So the prompt has your AI agent check first and fall back to installing from source if needed. Running only the free parts first shows you which items are "unknown," so you can decide whether they're worth spending credits on.

Step 2: Get the report and read it section by section

• Ads: Start with the ads that have run the longest and are still live. Note the pain points and promises they repeat, plus any offers. Then click through to see which page users land on.

• Social: Don't fixate on one viral post. Look at normal performance over a recent stretch and how often they post. Work out what they talk about, how they open their posts, and how the product shows up.

• Search pages: Check which kinds of pages are in their sitemap, such as "A vs. B" comparisons, "alternatives to X" pages, template pages, and integration pages. More pages doesn't mean more traffic. It only shows which search terms they're going after.

• Website redesigns: Flip through historical snapshots of their homepage. How did the main headline and button copy change? Did the target audience shift? Did the pitch move from features to results?

Two things are easy to misread. Having a platform's tracking code installed only shows they set it up, not that they're spending on it right now. An ad that has run for a long time is worth studying, but on its own it doesn't prove the ad makes money.

Step 3: Turn one finding into next week's experiment

Once you've read the report, the most important step is to test just one thing. Write each finding as evidence → judgment → action → metric. The prompt below has your AI agent help you organize them:

Why it's written this way: testing one thing at a time means that next week you can tell which change actually made the difference. The four parts, evidence → judgment → action → metric, force every idea to tie back to visible evidence and a number you can measure. For example, if you keep seeing a competitor post "before and after" comparison videos, you could make a similar short for your own product, then track how many people reach the landing page and how many start a trial.

Step 4: Re-run a week later and see what they changed

Run the same competitor again a week later with the same settings, then have your AI agent compare the two runs:

Use REM to run the same check on [competitor website] with last week's settings, then run the comparison (rem diff [competitor domain]).
In plain language, tell me: which new ads launched, which ones stopped, how much their social following grew, which new pages appeared on the website, and whether the homepage messaging changed.
First check whether both runs covered the same scope (for example, if I added a key this week that I didn't have last week). Flag any differences caused by a change in scope separately, and don't count them as changes on the competitor's side.
Finally, based on the result of the experiment I ran last week, suggest what I should test next week.

Why it's written this way: the author built a comparison command that lists new and stopped ads, follower growth, website page growth, and homepage positioning changes. But if the two runs used different settings, the extra items might just reflect a more complete search this time. That's why the prompt rules this out first.

Caveats

• Spot-check a few raw links: Does the social account the report identified really belong to this company? Is that ad still running? What date does the data run through? Open them yourself to confirm.

• Public data only: Don't log into their systems, get around paywalls, or use this to investigate specific individuals.

• Borrow tactics, not materials: Their ad videos, images, and copy are copyrighted.

• Enter keys yourself: Enter data service keys in the tool's settings yourself. Don't paste them into the chat.

References

• reverse-engineer-marketing (GitHub repo, MIT license)

• The sample report that ships with the project

• Scrape Creators